Sentence Transformers Embedding Provider

The Sentence Transformers embedding provider implements the embedding.Embedder interface using the HuggingFace Inference API’s feature-extraction pipeline. It supports any Sentence Transformers model hosted on HuggingFace, including the popular all-MiniLM, all-mpnet, and BGE model families.

Choose Sentence Transformers when you need access to the widest selection of community embedding models or when you want to use BGE models that rank competitively on MTEB benchmarks. This provider can also target a self-hosted HuggingFace Text Embeddings Inference (TEI) server for full control over hosting.

Installation

go get github.com/lookatitude/beluga-ai/rag/embedding/providers/sentence_transformers

Quick Start

package main

import (
    "context"
    "fmt"
    "log"
    "os"

    "github.com/lookatitude/beluga-ai/config"
    "github.com/lookatitude/beluga-ai/rag/embedding"
    _ "github.com/lookatitude/beluga-ai/rag/embedding/providers/sentence_transformers"
)

func main() {
    emb, err := embedding.New("sentence_transformers", config.ProviderConfig{
        APIKey: os.Getenv("HUGGINGFACE_API_KEY"),
    })
    if err != nil {
        log.Fatal(err)
    }

    ctx := context.Background()
    vec, err := emb.EmbedSingle(ctx, "Beluga AI is a Go framework for agentic systems")
    if err != nil {
        log.Fatal(err)
    }

    fmt.Printf("Vector length: %d\n", len(vec))
    fmt.Printf("Dimensions: %d\n", emb.Dimensions())
}

Configuration

Parameter	Type	Default	Description
APIKey	string	(required)	HuggingFace API token
Model	string	sentence-transformers/all-MiniLM-L6-v2	HuggingFace model ID
BaseURL	string	https://api-inference.huggingface.co	Inference API base URL
Timeout	time.Duration	0 (no timeout)	Request timeout
Options["dimensions"]	float64	Model-dependent	Override vector dimensions

Supported Models

Model	Default Dimensions
sentence-transformers/all-MiniLM-L6-v2	384
sentence-transformers/all-MiniLM-L12-v2	384
sentence-transformers/all-mpnet-base-v2	768
sentence-transformers/paraphrase-MiniLM-L6-v2	384
BAAI/bge-small-en-v1.5	384
BAAI/bge-base-en-v1.5	768
BAAI/bge-large-en-v1.5	1024

Any model on HuggingFace that supports the feature-extraction pipeline can be used.

Direct Construction

import (
    stemb "github.com/lookatitude/beluga-ai/rag/embedding/providers/sentence_transformers"
)

emb, err := stemb.New(config.ProviderConfig{
    APIKey: os.Getenv("HUGGINGFACE_API_KEY"),
    Model:  "BAAI/bge-large-en-v1.5",
})
if err != nil {
    log.Fatal(err)
}

Using BGE Models

The BGE (BAAI General Embedding) family provides competitive quality at different size tradeoffs:

// Small model -- fast, 384 dimensions
emb, err := embedding.New("sentence_transformers", config.ProviderConfig{
    APIKey: os.Getenv("HUGGINGFACE_API_KEY"),
    Model:  "BAAI/bge-small-en-v1.5",
})

// Large model -- higher quality, 1024 dimensions
emb, err := embedding.New("sentence_transformers", config.ProviderConfig{
    APIKey: os.Getenv("HUGGINGFACE_API_KEY"),
    Model:  "BAAI/bge-large-en-v1.5",
})

Self-Hosted Inference

Use the BaseURL parameter to point to a self-hosted HuggingFace Text Embeddings Inference (TEI) server:

emb, err := embedding.New("sentence_transformers", config.ProviderConfig{
    APIKey:  "unused",
    BaseURL: "http://localhost:8080",
    Model:   "sentence-transformers/all-MiniLM-L6-v2",
})